On Tue, 5 May 2020 at 18:05, Jerome Brunet <jbrunet@xxxxxxxxxxxx> wrote: > > > On Tue 05 May 2020 at 10:17, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote: > > > [...] > > > >> >> > + > >> >> > + return devm_of_clk_add_hw_provider(dev, of_clk_hw_onecell_get, > >> >> > + onecell_data); > >> >> > >> >> I think registering a provider for a module that does not provide clocks > >> >> to any other device is a bit overkill. > >> >> > >> >> I understand the matter is getting the per-user clk* pointer. > >> >> Since this is the module registering the clock, you can use clk_hw->clk > >> >> to get it. > >> >> > >> >> Once you have the clk* of the leaf clocks, you don't even need to keep > >> >> track of the clk_hw* since you are using devm_ > >> >> > >> >> Afterward, we should propably discuss with Stephen if something should > >> >> be added in CCF to get a struct clk* from struct clk_hw*. > >> >> > >> > > >> > [...] > >> > > >> > Hmm. > >> > > >> > I am not sure the above is a good idea, at all. Unless, I am > >> > misunderstanding your point, which may be the case. > >> > > >> > I think above "shortcuts" could lead to abuse of the clock framework > >> > and its internal data structures. When going forward, this could make > >> > it unnecessary harder to maintain the clock framework. > >> > > >> > I know, it's not my responsibility, but from my experience with MMC > >> > and SDIO interfaces, is that those have been too easy abuse - since > >> > most of the data structures and interfaces have been exported. Now, > >> > it's hard to roll back that, if you see what I mean. > >> > >> Indeed, it worth clarifying this first. > >> > >> With clk_register deprecated in favor of clk_hw_register, we are likely > >> to see that case rise elsewhere. > >> > > > > So, according to the separate discussion [1], I think we can let > > Martin decide what option to implement at this point. > > > > 1. Implement the "clk_hw_get_clk()" approach. The preferred option, > > but requires wider changes of the clock subsystem as well. > > > > 2. Keep the existing approach, with devm_clk_get(). I am fine with > > this as well, we can always switch to 1) later on. > > I have a problem with this approach. > The dt-bindings would include "#clock-cells = <1>" for a device that > does not actually provide and only needs it has a temporary work around. > Those bindings are supposed to be stable ... I agree, the bindings need to be stable. What is the problem of keeping "#clock-cells = <1>" around, when we move to a clk_hw_get_clk() approach in the next step? > > I have proposed 2 other short term solutions, let's see how it goes Yes, seems like we need to wait for Stephen's input then. Kind regards Uffe
